Learning Causal Networks from Episodic Data

Abstract. Discovering causal networks, especially from observational data alone, is a fundamental yet challenging task. Existing causal discovery algorithms not only rely on strict assumptions such as having i.i.d data, but are also limited to working with static, fully-specified datasets, rendering them incapable of learning causal networks in a continual fashion. In this short paper, we propose an information-theoretic approach that can learn causal networks in a continual fashion, does not require the i.i.d assumption on continually arriving data, and converges to the true underlying causal network as samples within the accumulated batches of data converge to the underlying data generating distribution. Our proposed approach, Continent, leverages the Algorithmic Markov Condition, a postulate by Janzing and Schoelkopf (2010), to discover causal networks in an online fashion. Continent is not only capable of continual learning, it also provides multiple plausible causal graphs at the end of each iteration, while the existing approaches can only predict a single causal network.


the Python source code (May 2024) by Osman Ali Mian.

Related Publications

Mian, O, Mameche, S & Vreeken, J Learning Causal Networks from Episodic Data. In: Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), ACM, 2024. (20% acceptance rate)
Mian, OA & Mameche, S An Information Theoretic Framework for Continual Learning of Causal Networks. In: Proceedings of the AAAI 2024 Continual Causality Bridge Program, PMLR, 2024.